Bayesian methods for battery state of health estimation

Thesis

Estimating the state of health of battery energy storage systems is key to their operational safety and reliability, both of which affect lifetime cost. However, accurate estimation of state of health remains challenging, as measurement techniques used in laboratory environments are not available in real-world operating environments. In this work, a framework is developed that combines the relative strengths of commonly applied model- and data-driven approaches to state of health estimation. Gaussian process regression, a flexible Bayesian method of learning arbitrary functions from input-output data, is applied to estimate the parameters of low-order battery models as functions of internal states, operating conditions and lifetime. The approach is first motivated by the difficulty of parameter identification for physics-based battery models from real-world data. Then it is shown how electrical equivalent circuit models can be extended to include parameter dependencies on operating conditions and lifetime in a data-driven manner. The framework is then applied to two different usage scenarios. First, internal resistance is estimated for a fleet of solar-connected lead-acid batteries located in sub-Saharan Africa, where the resulting health metric is shown to provide an early indication of end-of-life failure. Second, a first-order RC circuit, coupled with a one-state thermal model, is parameterised in a joint process that also simultaneously estimates battery states, using data from a Li-ion cell under laboratory conditions. The only prerequisite was the cell-level open-circuit voltage versus charge curve and, in the case of the Li-ion cell model, a single thermal parameter. Given this, the method is agnostic to chemistry and battery construction. Enabling robust and fast state of health estimation for large fleets of batteries has the potential to `close the loop' in terms of battery energy storage system design. Instead of performing laboratory-based ageing experiments, field data can be used directly to determine factors that affect battery life. By incorporating this information into fault diagnosis and health-aware battery management systems, safety and reliability will be improved. Furthermore, with deeper understanding of degradation in the real world, better design of energy storage systems will ultimately lead to better cost efficiency through reduced over-engineering.

Authors: Aitio, A

• DOI: 10.5287/ora-r51v7bvdr
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: lead-acid; Gaussian Process regression; battery state of health; parameterisation; li-ion; bayesian methods
• Subjects: Machine learning; Battery management systems; Energy storage